DIAGNOSTIC PERFORMANCE AND ACCURACY OF THREE AI MODALITIES IN ASSESSING THE MANDIBULAR CANAL RELATIONSHIP

European Congress of DentoMaxilloFacial Radiology, Iasi, Romanya, 25 - 27 Haziran 2026, ss.97-98, (Özet Bildiri)

Aim: The aim of this study was to evaluate the diagnostic efficacy of Gemini AI’s three different

modes in assessing the anatomical relationship between mandibular third molars and the

mandibular canal, utilizing Cone-Beam Computed Tomography (CBCT) as the reference

standard.

Materials and Methods: An oral radiologist retrospectively evaluated existing CBCT scans of

116 mandibular third molars from 59 patients to determine their relationship with the mandibular

canal as the gold standard. Subsequently, the corresponding panoramic radiographs of these

patients were introduced to the Gemini AI system. The diagnostic performance of three different

AI modes (Gemini Think, Gemini Flash and Gemini Pro) was then assessed by comparing their

outputs against the CBCT findings.

Results: The Gemini AI modes exhibited varying tooth detection rates, with failures occurring

in 14.6% to 20.7% of cases. Gemini Pro demonstrated superior sensitivity (98.25%) but

significantly compromised specificity (7.14%), indicating a high rate of false positive

assessments. Conversely, Gemini Flash yielded the highest specificity (58.97%) but displayed

suboptimal sensitivity (37.93%). Overall diagnostic accuracy remained moderate, ranging from

46.39% to 60.87%. Concordance with the gold standard was marginal, with Kappa values (k= -

0,029 to 0,134) indicating slight to poor agreement.

Conclusion: Although Gemini AI modes show promise in identifying canal involvement,

substantial tooth detection failures and high false positive rates limit their current clinical

reliability. At present, these AI modes should be considered diagnostic adjuncts requiring expert

verification rather than substitutes for CBCT in preoperative surgical planning.